Challenges of predicting rainfall changes under global warming: back to fundamentals

Transcription

1 Challenges of predicting rainfall changes under global warming: back to fundamentals J. David Neelin 1 K. Hales 1, B. Langenbrunner 1, J. E. Meyerson 1, S. Sahany 1, B. Lintner 2, R. Neale 5, O. Peters 4, C. E. Holloway 3, M. Munnich 6, H. Su 7, C. Chou 8, J. McWilliams 1, A. Bracco 9, H. Luo 9 Examples and issues with precipitation simulation: global warming, El Niño Parameter sensitivity/optimization---implications The onset of strong convection: constraining climate model representations Outlook 1 Dept. of Atmos. and Ocn. Sci, U.C.L.A., 2 Rutgers Univ., 3 Univ. Reading, 4 Imperial College London, 5 Nat. Ctr. Atm. Res, 6 Swiss Fed. Inst. Tech., 7 JPL, 8 Academia Sinica, 9 Georgia Inst. Tech

2 Global warming scenarios: IPCC* 2007 & ~2013 Inputs for Coupled Model Inter-Comparison Project (CMIP) 3 & 5 for *Intergovernmental Panel on Climate Change Assessment Reports 4 & 5 Historical period est. observed greenhouse gas + aerosol forcing, followed by Representative Concentration Pathway (RCP)

3 Global warming as simulated in climate models ~2007 Global avg. sfc. air temp. change (ann. means rel. to base period) Greenhouse gas + aerosol forcing: Est. observed, followed by SRES A2 * scenario (inset) in 21 st century T s (C) *SRES: Special Report on Emissions Scenarios A2: uneven regional economic growth, high income toward non-fossil, population 15 billion in 2100; similar to an earlier business-as-usual scenario IS92a

4 Global warming as simulated in climate models CMIP5 Global avg. sfc. air temp. change (ann. means rel. to base period) Greenhouse gas + aerosol forcing: Est. observed followed by RCP8.5 from 2005 T s (C) Global average temperatures *Representative Concentration Pathway specified: not full Earth System Model, i.e., carbon cycle feedbacks etc. not active in runs shown here

5 Surface air temperature change for three models * annual avg. (rel. to ) CMIP5 NCAR- CCSM4 IPSL- CM5A-R MRI- CGCM3 *Unexplained acronyms denote climate model names

6 Examples and issues with precipitation simulation: global warming, El Niño Severe problems with model disagreement on precipitation change at regional/seasonal scales, markedly so in tropics some agreement on large-scale or amplitude Poor simulation of El Niño remote precipitation anomalies Sensitivity to differences in model parameterizations Teleconnections of errors in other parts of the climate system to influence edges of convection zones/storm tracks e.g., IPCC 2001, 2007; Wetherald & Manabe 2002; Trenberth et al 2003; Neelin et al. 2003; Maloney and Hartmann 2001; Joseph and Nigam 2006; Biasutti et al. 2006; Dai 2006; Tost et al. 2006; Bretherton 2007, Frierson,...

7 Precipitation: climatology (CMAP * : ) January Note intense tropical moist convection zones (intertropical convergence zones) July Later: 4 mm/day contour as indicator of precip. climatology *CPC Merged Analysis of Precipitation (CMAP)

8 Neelin, 2011,Climate Change and Climate Modeling Cambridge UP Observed (CMAP) and CMIP3 coupled models 4 mm/day precip. contour Coupled simulation climatology (20 th century run, ) December-February precipitation climatology June - August precipitation climatology CPC Merged Analysis of Precipitation (CMAP)

9 Observed (CMAP) and CMIP5 coupled models 4 mm/day precip. contour Coupled simulation climatology (20 th century run, ) December-February precipitation climatology June - August precipitation climatology Coupled Model Intercomparison Project (CMIP5) Analysis: J. Meyerson

10 IPCC 2007 multi-model, annual mean precipitation change ( relative to ) High latitudes wetter Subtropics dryer/expand Deep tropics wetter Stippled where 80% of the models agree on sign of the mean change. Note typical magnitudes <0.5mm/d. IPCC 4th Assessment Report (WG1 2007, chpt 10; A1B Scenario)

11 Fourth Assessment report models Data archive at Lawrence Livermore National Labs, Program on Model Diagnostics and Intercomparison SRES A2 scenario (heterogeneous world, growing population, ) for greenhouse gases, aerosol forcing Neelin, Munnich, Su, Meyerson and Holloway, 2006, PNAS Precipitation change: HadCM3, Dec.-Feb., avg minus avg. 4 mm/day model climatology black contour for reference mm/day CMIP3

12 NCAR_CCSM3 JJA Prec. Anom. CMIP3

13 CCCMA JJA Prec. Anom. CMIP3

14 CNRM_CM3 JJA Prec. Anom. CMIP3

15 CSIRO_MK3 JJA Prec. Anom. CMIP3

16 GFDL_CM2.0 JJA Prec. Anom. CMIP3

17 GFDL_CM2.1 JJA Prec. Anom. CMIP3

18 UKMO_HadCM3 JJA Prec. Anom. CMIP3

19 MIROC_3.2 JJA Prec. Anom. CMIP3

20 MRI_CGCM2 JJA Prec. Anom. CMIP3

21 NCAR_PCM1 JJA Prec. Anom. CMIP3

22 MPI_ECHAM5 JJA Prec. Anom. CMIP3

23 CMIP5/IPCC 5th Assessment report models Representative Concentration Pathway RCP 8.5 (akin to CMIP3 A2 scenario) for greenhouse gases, aerosol forcing Precipitation change: HadCM3, Dec.-Feb., avg minus avg. 4 mm/day model climatology black contour for reference Analysis: J. Meyerson NCAR Community Climate System Model mm/day CMIP5

24 BCC-ESM1-1 JJA Prec. Anom. Beijing Climate Center, China CMIP5

25 CanESM2 JJA Prec. Anom. Canadian Center for Climate Modelling and Analysis, Canada. CMIP5

26 CCSM4 JJA Prec. Anom. NCAR Community Climate System Model CMIP5

27 CNRM-CM5 JJA Prec. Anom. Centre National de Recherches Mereorologiques/ Centre Europeen de Recherche et Formation Avancees en Calcul Scientifique, France. CMIP5

28 CSIRO-MK3 JJA Prec. Anom. Commonwealth Scientific and Industrial Research Organization, Aus. CMIP5

29 GISS-E2-R JJA Prec. Anom. Goddard Institute for Space Studies CMIP5

30 INMCM4 JJA Prec. Anom. Institute for Numerical Mathematics, Russia. CMIP5

31 IPSL-CM5A JJA Prec. Anom. Institut Pierre Simon Laplace, France. CMIP5

32 MRI-CGCM3 JJA Prec. Anom. Meteorological Research Institute, Japan CMIP5

33 NORESM1-M JJA Prec. Anom. Norwegian Climate Center, Norway CMIP5

34 Mechanisms & constraints from moisture/energy budgets Moisture budget for perturbations P ' = <q ' `v > <`v q ' > Precip Rich-get-Richer Upped-ante <`q v ' > Convergence Fb + E ' + Evap 0. At global scale neglect transport P ' E ', set by surface energy balance small increase (e.g., Allen & Ingram 2002, ) 0.1 Warmer temperatures & Clausius-Clapeyron q ' tends to increase [Interplay with convection and dynamics q ' ] < >= vertical average; q ' specific humidity; ' denotes changes

35 Mechanisms & constraints from moisture/energy budgets Moisture budget for perturbations P ' = <q ' `v > <`v q ' > Precip Rich-get-Richer Upped-ante <`q v ' > Convergence Fb + E ' + Evap Rich-get-richer mechanism* Subtropics: low-level divergence so q ' increase Precip decrease Convergence zones: vice versa Convergence zones Subtropics *(a.k.a. thermodynamic component):

36 The Rich-get-richer mechanism Center of convergence zone: incr. moisture convergence incr. precip Descent region: incr. moisture divergence; less often meets conv. threshold Chou & Neelin, 2004, Held & Soden 2006, Chou et al 2008

37 Mechanisms & constraints from moisture/energy budgets Moisture budget for perturbations P ' = <q ' `v > <`v q ' > Precip Rich-get-Richer <`q v ' > Convergence Fb Upped-ante [Regional differences] + E ' + Evap a. Atm. energy budget to approx. v ' (Chou & Neelin 2004) b. Neglect v ', (Held and Soden 2006; plausible for large scales) v ' large at regional scales! a major factor in uncertainty v q ' in particular regions Averaging over larger scales, e.g., latitude bands; or a an ensemble of models that disagreed on location of strong convergence change can reduce the visibility of the convergence feedback terms

38 West Coast rainfall change under global warming DJF Prec. Anom. ( )- ( ), RCP 8.5 scenario Analysis: J. Meyerson CMIP5

39 CMIP5 CanESM2 DJF Prec. Anom.

40 CMIP5 CCSM4 DJF Prec. Anom.

41 CMIP5 CNRM-CM5 DJF Prec. Anom.

42 CMIP5 CSIRO-MK3 DJF Prec. Anom.

43 CMIP5 GISS-E2-R DJF Prec. Anom.

44 CMIP5 INMCM4 DJF Prec. Anom.

45 CMIP5 IPSL-CM5A DJF Prec. Anom.

46 CMIP5 MRI-CGCM3 DJF Prec. Anom.

47 CMIP5 NORESM1-m DJF Prec. Anom.

48 How do the models do for El Niño/Southern Oscillation (ENSO)? A phenomenon we can observe Important for interannual prediction Satellite precipitation retrievals since 1979 Atmospheric model component runs with observed sea surface temperature (SST) or ocean atmosphere models Rank correlation/regression/compositing of events based on an equatorial Eastern Pacific SST index Nino3.4

49 ENSO teleconnections to regional precip. anomalies Tropospheric temperature anomaly Su & Neelin, 2002 See Newell and Weare (1976); Salby & Garcia 1987; Yulaeva & Wallace (1994); Wallace et al. (1998); Chiang and Sobel (2002); Kumar & Hoerling (2003); Su and Neelin 2003; Sperber and Palmer 1996, Giannini et al 2001; Saravanan & Chang, 2000; Joseph & Nigam 2006,

50 Observed Nino3.4 rank correlations (Dec.-Feb.) CMAP CPC Merged Analysis of Precipitation Compare to preliminary results from CMIP5 models Analysis: B. Langenbrunner

51 CMIP5 models nino3.4 rank corr. AMIP runs(dec.-feb.) CanAM4 Canadian Center for Climate Modelling and Analysis, Canada. CMIP5

52 CMIP5 models nino3.4 rank corr. AMIP runs(dec.-feb.) CCSM4 NCAR Community Climate System Model CMIP5

53 CMIP5 models nino3.4 rank corr. AMIP runs(dec.-feb.) CNRM Centre National de Recherches Mereorologiques/ Centre Europeen de Recherche et Formation Avancees en Calcul Scientifique, France. CMIP5

54 CMIP5 models nino3.4 rank corr. AMIP runs(dec.-feb.) CSIRO Commonwealth Scientific and Industrial Research Organization, Aus. CMIP5

55 CMIP5 models nino3.4 rank corr. AMIP runs(dec.-feb.) HadGEM-A Met Office Hadley Centre, UK. CMIP5

56 CMIP5 models nino3.4 rank corr. AMIP runs(dec.-feb.) INMCM4 Institute for Numerical Mathematics, Russia. CMIP5

57 CMIP5 models nino3.4 rank corr. AMIP runs(dec.-feb.) IPSL Institut Pierre Simon Laplace, France. CMIP5

58 CMIP5 models nino3.4 rank corr. AMIP runs(dec.-feb.) MPI Max Plank Institute, Germany CMIP5

59 CMIP5 models nino3.4 rank corr. AMIP runs(dec.-feb.) MRI Meteorological Research Institute, Japan CMIP5

60 CMIP5 models nino3.4 rank corr. AMIP runs(dec.-feb.) NorESM1-m Norwegian Climate Center, Norway Analysis: B. Langenbrunner CMIP5

61 What is being done across the field? Higher-resolution models (no guarantee) Regional models (boundary conditions from global models) Multimodel ensemble means and general (vs. regional) statements Large satellite data sets, field campaigns, monitoring at Atmospheric Radiation Measurement sites. Need to digest in ways that better constrain parameterizations* of moist convection at short time scales Understanding of parameter sensitivity/uncertainty quantification; practical means of optimizing models with available data Alternatives to point by point multi-model ensemble mean *Parameterization: representation of bulk effects of small-scale phenomenon as a function of grid-scale variables

62 Hypothesis for disagreement on regional scale: models have similar processes for precip increases and decreases but the geographic location is sensitive to differences in model clim. of wind, precip; convective closure (e.g. threshold) agreement on amplitude measure* suggests strong regional changes are likely that are not reflected in multi-model averages. *e.g., spatial projection of precip change on each model s own characteristic pattern

63 Despite disagreement on precise location, seek measures of extent of precip change that are more predictable E.g., amplitude of precip incr/decr pattern shows better agreement Projection of Jun- Aug (30yr running mean) precip pattern onto normalized positive & negative latecentury pattern for each model Neelin, Munnich, Su, Meyerson and Holloway, 2006, PNAS CMIP3

64 Integrated measures of regional precip. change cont d Fraction of the globe with annual precipitation that would be in highest 5% (~20 year wet spell) during base period ( ) for each model Analysis: B. Langenbrunner ; five year running mean shown for graphical clarity CMIP5

65 Integrated measures of regional precip. change cont d Fraction of the globe with annual precipitation that would be in lowest 5% (~20 year drought) during base period ( ) for each model Analysis: B. Langenbrunner ; five year running mean shown for graphical clarity CMIP5

66 Are there fundamental considerations in climate model sensitivity that techniques borrowed from optimization methods can help with? Precipitation parameter sensitivity a critical limitation to confidence levels in regional scale projections---arguably more important for impacts this century than climate sensitivity for global average temperature How nonlinear is this sensitivity? E.g., convection has sharp threshold for onset, but climate avgs over many instances Can we infer implications for the model improvement process and the use of multi-model ensemble averages to estimate projected precipitation changes? Neelin, Bracco, Luo, McWilliams, Meyerson, 2010, PNAS.

67 Precipitation sensitivity cont d Interest in systematic parameter sensitivity (esp. global avg climate sensitivity) and optimization in climate models (Severijns & Hazeleger 2005 Clim. Dyn., Stainforth et al Nat., Jones et al Clim. Dyn., Knight et al PNAS, Kunz et al Clim. Dyn., Jackson et al J. Clim., Rougier et al J. Clim., ) # parameters N can easily be >10; a priori feasible range Brute force sampling at density s gives order s N problem, but e.g. ~N 2 depending on nature of parameter dependence. Rough/smooth? High-order nonlin? Irreducible imprecision? Here examined in the ICTP climate model *International Centre for Theoretical Physics atmospheric general circulation model: ICTP AGCM; Molteni F., 2003, Climate Dyn.; Bracco et al. 2004, Climate Dyn.) Eight Sigma-levels, spectral triangular truncation T30 ~3.75 x 3.75-degree

68 Parameter dependence of RMS error* of June-Aug. precip as a function of cloud albedo, convective rel. hum., RH conv AGCM ensemble mean over 10*25-year runs, (with observed sea surface temp.). Vertical size of symbol=2*standard error of ensemble mean Individual ensemble members shown for RH conv Neelin, Bracco, Luo, McWilliams, Meyerson, 2010, PNAS. *(vs. NCEP reanalysis)

69 Metamodel fit to param. dependence of AGCM fields Try quadratic metamodel on space of N parameters μ i for field j N N N j j a b std i i ij i j i i 1 j 1 Simple but important: linear coefficient a i (x,t) & quadratic coefficient b ij (x,t) are spatial & seasonal fields e.g. of entry-level strategy for computationally-expensive blackbox functions (cf. review by Shan & Wang, 2010, Struct. Multidisc. Optim.) j can be a climatological field, anomaly regression, or other statistic from model output. Adopt multi-objective approach (for each field). 1 / 2 2 Then construct objective function, e.g., rms error j j i (or sq. error, spatial correlation ) with typically a spatial mean, j obs observed, j std the GCM for standard parameters First fit: a i (x,t), b ii (x,t) from the 2N endpoints of the μ i ranges (order N integrations even if add redundant points). For off-diagonal b ij =b ji : order N 2 (at least N(N-1)/2 simulations).

70 RMS error of June-Aug. precipitation (vs. NCEP) as a function of cloud albedo, convective RH AGCM ensemble average versus linear and quadratic metamodels. Note negative curvature for relative humidity, due to ~quadratic nonlinearity in spatial field. No interior minimum boundary solution in constrained optimization problem Neelin et al

71 RMS error of June-Aug. precipitation (vs. reanalysis) as a function of convective RH but AGCM coupled to a mixed-layer (ML) ocean (preindustrial CO 2 ) Same properties in coupled model AGCM-ML average (250 yrs) versus linear and quadratic metamodel. Negative curvature for relative humidity (assoc. with ~quadratic nonlinearity in param. dependence of spatial fields) as in specified SST case. Vertical size of symbol=2*standard error Neelin, Bracco, Luo, McWilliams, Meyerson 2010, PNAS.

72 *case of interior minimum for diagonally dominant Hessian A Role of high dimensional fields in improvement challenges Common experience: One region improves but another gets worse! Illustrate with case* of objective function f, (e.g. RMS precip error) with standard case error j err μi f = g i + A ii μ i = 0 g i = 2 a i j err, A ii = 2( a 2 i +2 b ii j ) spatial average, metamodel linear coeff a i, quadratic b ij. For simplicity neglect b ij in curvature. μ i = - a i j err / a 2 i If sensitivity a i had same spatial pattern as the standard case error j err j std j obs this would cancel the error. Instead, compromise between reducing j err and introducing new error.

73 Parameter dependence for precipitation (etc) changes under global warming: Implications for multi-model ensemble average Does sensitivity across the feasible parameter domain provide a prototype for differences among models? If so, multi-model ensemble average ~ random sampling If parameter dependence is linear, and distribution of sample points is unbiased with respect to true parameter value multi-model ensemble average should work well parameter directions with (1) strong nonlinearity or (2) boundary optima (suggesting sampling across feasible range likely biased) can limit usefulness of multi-model ensemble average; e.g., convective rel. humidity param.

74 RMS difference (vs. Rh conv =0.9) of June-Aug. precipitation change as a function of convective RH for AGCM-ML 2xCO 2 minus preindustrial CO 2 AGCM ensemble average versus linear and quadratic fit. Note negative curvature for relative humidity, due to quadratic effects. Global warming precipitation change parameter dependence Neelin, Bracco, Luo, McWilliams, Meyerson 2010, PNAS.

75 Global warming precipitation change parameter sensitivity Ensemble-mean JJA precipitation (as a departure from the annual mean) for Conv. rel. hum. param max relative to the standard case for AGCM coupled to a mixedlayer ocean: change for 2xCO 2 minus preindustrial. Linear contribution Nonlinear contribution Neelin, Bracco, Luo, McWilliams, Meyerson 2010, PNAS.

76 Implications for multi-model ensemble average

77 Back to fundamentals: better constraining and representing processes at small time/space scales Column integrated water vapor observational estimate from microwave retrievals* *Satellite instruments: AMSR-E, SSMI; dynamic interpolation Wimmers & Velden (2007); footprint of input ~15 km; swath width ~1400 km; retrieval algorithm Alishouse et al. (1990)

78 Column water vapor from NCAR CAM4* at resolution *National Center for Atmospheric Research Community Atmosphere Model, HOMME spectral element dynamical core. Courtesy Mark Taylor (Sandia NL) & Rich Neale (NCAR).

79 An example of quantifying convective onset: Precipitation binned by column water vapor, w buoyancy & precip. pickup at high w Entraining convective available potential energy (CAPE) can match onset---if include enough turbulent entrainment into convecting parcel w useful because lots of microwave data available Neelin, Peters, Lin, Holloway & Hales, 2008, Phil Trans. Roy. Soc. A

80 Transition to strong convection: Precip. dependence on tropospheric temperature & column water vapor Averages conditioned on vert. avg. temp. ^ T, as well as w (T mb from ERA40 reanalysis) Power law fits above critical: w c changes, same [note more data points at 270, 271] Analysed in tropics 20N-20S E. Pacific Neelin, Peters & Hales, 2009 JAS

81 Collapsed statistics for observed precipitation Precip. mean & variance dependence on w normalized by critical value w c ; occurrence probability for precipitating points (for 4 T values); Event size distribution at Nauru

82 Tropospheric temperature T (k) ^ T E. Pacific x p

83 Tropospheric temperature T (k) Defines an empirical thermodynamic surface for the onset of strong convection to test models Not a constant fraction of column saturation ^ T E. Pacific x p

84 Transition to strong convection: High-resolution global model (CAM3.5, 0.5 ) compared to observations (TMI) Model Obs Sahany et al. 2011, subm.

85 Transition to strong convection: High-resolution global model (CAM3.5, 0.5 ) compared to observations (TMI) Obs Model Convective onset boundary Sahany et al. 2011, subm.

86 Transition to strong convection: Obs. & model compared to simple convective plume instability calculation with different entrainment assumptions Obs CAM3.5 entrainment Low entrainment Low values of entrainment are inconsistent with observed onset

87 Transition to strong convection: simulation of current conditions Community Climate System Model 4 (CAM4, 1 ) Historical run Column water vapor w (mm) CAM4 Instantaneous precipitation data: R. Neale, Analysis K. Hales Conditionally avg. Precip P for bins of Tropospheric bulk temperature T (K)

88 Transition to strong convection: simulation under global warming Community Climate System Model 4 (CAM4, 1 ) Representative Concentration Pathway run RCP Conditionally avg. Precip P for bins of Tropospheric bulk temperature T (K) Column water vapor w (mm) CAM4 Instantaneous precipitation data: R. Neale, Analysis K. Hales

89 Importance of very small scales Importance of entrainment to the onset of deep convection Explains the high sensitivity to free tropospheric water vapor (above the boundary layer) Bad news: Beyond the resolution of global climate models anytime soon (100m vs. 100 km) Good news: work for cloud resolving modelers; new observations add constraints; revised model comes close Bad news: interacts with other poorly constrained small scale processes cloud microphysics Kirshbaum 2011

90 Outlook The regional scale changes in the hydrological cycle are arguably the most important aspect of climate sensitivity over the 21 st century Reducing regional uncertainty remains challenging with the current set of CMIP5 models---regions of agreement TBD Using climate model precipitation projections: Caution on simple statements; measure of uncertainty on multi-model ensemble mean; specific model validation for key phenomenon in the region of interest for each member of the ensemble

91 Outlook The regional scale changes in the hydrological cycle are arguably the most important Will we do any better at reducing uncertainty? Current tackling of small scale processes, scale interactions, new observational constraints, systematic parameter estimation methods, seem likely to yield progress--- although not high precision by July 2012

92 Some connections Long tails seen in the probability distribution of water vapor also occur for chemical tracers including CO2: (B. Lintner, B. Tian, Q. Li, L. Zhang, P. Patra, M. Chahine) And surface temperature (T. Ruff) Simple stochastic model Fokker-Planck solutions indicate processes (S. Stechmann) Nastier parameter dependence can occur (M. Chekroun et al.) Do constraints on entrainment combine with new proxy data to resolve a surface temperature vs. glacial elevation conundrum at last glacial maximum? (A. Tripati, S. Sahany, D. Pittmann, R. Eagle, J. Eiler, J. Mitchell, L. Beaufort) theory for inflow air mass interacting with convective onset at the margins of convection zones can be tested in models (H.Y. Ma, C.R. Mechoso, X. Ji)